Paste this into any Claude Code session: /context. Before you've typed a single word of your actual task, you'll see something like 45,000 tokens already spent: system prompt, tool schemas, MCP definitions, CLAUDE.md, memory files. That's your startup overhead, and you pay it again on every single turn, in dollars and in model attention.
Everyone I show this to asks the same two questions: which part is waste and what does it cost me. So I built a tool that answers both. Paste your /context output into the Claude Code Context Auditor and you get a category-by-category breakdown, flagged recommendations, and the idle spend in dollars per session and per month. Free, no sign-up, and 100% client-side: your context report never leaves your browser.
What it does
Three things, in the time it takes to paste:
1. Parses the breakdown. It reads the category lines from /context (MCP tools, memory files, custom agents, messages, free space) and renders them as a ranked bar chart, so the biggest line item is impossible to miss. In most sessions I audit, it's MCP tool schemas, and the owner is surprised every time.
2. Applies audit thresholds. Each category gets a verdict based on the heuristics I use in real engagements. MCP definitions over 25k tokens: high severity, audit your servers, anything that isn't live data should be a skill. Memory files over 8k: apply the residency test from the CLAUDE.md guide, because every always-loaded sentence is a per-session tax. Conversation history over 60k: your exploration belongs in subagents that return conclusions, not transcripts.
3. Prices it. Pick a model and how many turns a typical session runs, and it multiplies out what that startup overhead costs. The math is blunt: overhead tokens are re-billed as input on every turn. A 40k-token startup at Sonnet pricing across a 40-turn session is about $4.80 of pure overhead, per session, before any work happens. Run one heavy session per working day and that's roughly $100 a month for the privilege of loading MCP servers you forgot you installed. Opus pricing: multiply by five.
There's a "copy audit report" button that produces a plain-text summary, because the fastest way to get a teammate to trim their CLAUDE.md is to paste their own numbers at them.
Why a browser tool and not another CLI
There are good CLI analyzers and hook-based dashboards for this on GitHub. But the moment I want a teammate, a client, or a Reddit thread to see their waste, "install this CLI and pipe your output through it" loses to "paste it here." Zero-install is the whole point: the audience for this tool is precisely the people who haven't already built context-analysis tooling.
And because the input is just text you paste, there's nothing to trust. No upload, no account, no telemetry. The parsing and pricing run as JavaScript in your tab, same as the other 89 tools on this site.
The audit loop, one more time
The tool automates step one of the FinOps loop from the context engineering post: audit, right-size, schedule, isolate, monitor. The remaining steps are judgment calls, but they're fast once the numbers are in front of you:
- Audit: paste
/contextinto the auditor. Look at the top bar. - Right-size: demote by access pattern. Always-true goes in CLAUDE.md, procedures become skills, path-specific rules get scoped, fetchable docs stay on disk.
- Schedule: must-always-happen logic goes in hooks, not prompts.
- Isolate: expensive exploration goes to subagents.
- Monitor: re-paste after every cleanup, and again next month. Context bloat regrows exactly like cloud spend: one reasonable-sounding addition at a time.
Run your audit. If your MCP bar is the tallest one, you already know what to do.